#seq2seq
And today's AI products are the direct descendants of this nearly century of work. They didn't just emerge from the ether, there's a direct line from Hinton's backpropagation to word2vec to seq2seq to Transformers, which itself took years to mature.
October 25, 2026 at 9:04 PM
A new AI review! bentrevett/pytorch-seq2seq ⭐3.7/5.0
This repository is a well-known, educational set of PyTorch notebooks that walk through implementing core sequence-to-sequence (seq2seq) neural machine translation models (German→English), progres...
https://gitrated.com/bentrevett/pytorch-seq2seq
October 20, 2026 at 12:56 PM
音声合成の日本語化手法が一般化した話
漢字の読みの曖昧さがSeq2seq音声合成の学習を崩す仕組みと、アクセントを記号で入力に埋め込む日本語化の考え方を述べた記事です。

漢字は読みが定まらず音声との対応を崩すため、日本語では読みとアクセントを記号として入力に埋め込む方が有効です。
音声合成の日本語化手法が一般化した話
zenn.dev
October 13, 2026 at 3:47 PM
I worked with trying to predict secondary DNA structure in my masters a few years ago with different types of seq2seq NN models trained off basic energy calculations to see if it was even feasible and while it worked, it never got anywhere close to even 90% for the amount of computation it took.
September 26, 2026 at 4:08 AM
PyTorch自然言語処理プログラミング word2vec/LSTM/seq2seq/BERTで日本語テキスト解析! impress top gearシリーズ (新納浩幸) が、紀伊國屋電子書籍の【出版社合同】秋のIT書フェア2026で1540円引きの、1540円(50%OFF)+14ポイント還元になりました。10/8(木)まで。
PyTorch自然言語処理プログラミング word2vec/LSTM/seq2seq/BERTで日本語テキスト解析! impress top gearシリーズ
著者:新納浩幸(著) 出版:インプレス 2021/3/18(木)配信
5leaf.jp
September 25, 2026 at 9:25 PM
CDS-BART combines SentencePiece BPE tokenization with BART's denoising seq2seq architecture, pre-trained on 60 million CDSs from nine NCBI RefSeq taxonomic groups to handle sequences up to ~4 kb.
September 23, 2026 at 10:02 AM
Further, social media feeds since ~2020 have largely been using seq2seq transformers, i.e. exactly the same things as in LLMs.

Scrolling anywhere with "an algorithm" (X, IG, TT, YT) should be understood as "talking to an AI", the way AI is understood today, with all the same harms and risks.
September 19, 2026 at 3:14 AM
Jevに困惑している人たちは何なんだ
BERTとかseq2seqの概念に近いと思うんだが
September 17, 2026 at 10:55 AM
fact: not only is machine translation generative ai (seq2seq model), it's also one of the oldest consumer applications of the tech, starting long before the ai bubble was a thing
September 13, 2026 at 6:29 AM
Being totally real, I read the description of how seq2seq etc were able to do language translation via aligning vector spaces with basic transforms and I was like well shit that’s it then.
The weird thing for me is that I was basically convinced at the T5/GPT-3 level, that it was doing almost everything important, but did not trade any stock on that basis.
One thing that's fun about this moment is: pragmatic software people who have receipts to show that they were AI skeptics until 6-18 months ago.

By contrast, I am the least neutral observer possible. GOFAI dad / traumatized by AI winter / grew up hating Searle like the six-fingered man.
September 8, 2026 at 4:01 AM
"Excited after reading 'Attention Is All You Need'! Transformers revolutionize seq2seq modeling, ditching RNNs and CNNs for pure attention. Faster training,
superior performance. Positional encoding and multi-head attention make it a game-changer in NLP. Can't wait to see its impact! #NLP”
-qwq 32b
December 4, 2024 at 7:06 PM
I mean whatever semantic information they happen to collect and represent.

And yeah, what you describe isn't how natively multimodal LLMs work. You are probably thinking of the older seq2seq way.
July 26, 2025 at 9:00 PM
That is an inaccurate retelling of history. A lot of what modern LLMs are based on was originally created for machine language translation, not content creation. Look for seq2seq to see how this came together originally.
January 30, 2025 at 12:46 PM
Discover the different types of transformer models: Full Transformer Models with Encoder-Decoder Architecture, Encoder-Only Models, and Decoder-Only Models. Learn more about #transformerarchitecture #seq2seq. https://machinelearningmastery.com/encoders-and-decoders-in-transformer-models/
https://machinelearningmastery.com/encoders-and-decoders-in-transformer-models/
machinelearningmastery.com
June 3, 2025 at 11:05 AM
ShahZeb Khan, Chiara Pallara, Barbara Monti, Alexis Molina: De Novo Design of SIK3 Inhibitors via Feedback-Driven Fine-Tuning of Seq2Seq-VAE https://arxiv.org/abs/2511.06930 https://arxiv.org/pdf/2511.06930 https://arxiv.org/html/2511.06930
November 11, 2025 at 6:49 AM
Not all AI models are built the same 🤖

Encoders, decoders, and seq2seq models power everything from chatbot to translation.

Learn the basics in just 5 mins:
bit.ly/4i84CiV
How to Pick the Best AI Model for Your Use Case
Understanding the transformer architectures behind AI Models for chatbots, translation, and text generation
bit.ly
February 22, 2025 at 3:27 PM
このくらいが普通の高校生なのにローカルLLM会の高校生ときたら……Transformerレベル(アーキテクチャレベル)の理論なら、
Seq2Seq->ResNet->Transformer->S4・Hyena・Mamba
という順に詳しく解説して、実装していくのが堅そう。とか抜かす超人しかいねえ。
あかん、ど忘れがひどくなってきたな・・・
人やモノの名前がとっさに出てこない😅
April 19, 2024 at 4:44 AM
Transformerレベル(アーキテクチャレベル)の理論なら、
Seq2Seq->ResNet->Transformer->S4・Hyena・Mamba
がいいですねと入門会の議題。
GPT4まで解説用とすな・・・
April 11, 2024 at 1:20 AM
scaling inscrutable seq2seq system to de-inscrute inscrutable skeets ✅
May 21, 2025 at 2:37 AM
9. Compact seq2seq models improve ASR error correction efficiency LINK
10. Claude-video tool for automated video frame extraction and transcription LINK
July 7, 2026 at 1:00 PM
I'm surprised it's not setting off red flags for people, since seq2seq and GloVe predate Arrival.
December 1, 2024 at 8:08 PM
Even something as primitive as word2vec and seq2seq have a world model. To the point that you can go very literal with the concept of "world model" and recreate a map of the world just from plotted probes.
October 12, 2024 at 12:23 AM
World models with concepts of position and temporality have been amply demonstrated in models as simplistic as seq2seq (you can even graph out maps with probes), let alone modern Transformers with billions or even trillions of parameters. If the sentence it is tasked to complete starts out:
September 25, 2024 at 3:25 PM
Words do not exist in a vacuum. Words are a reflection of the world that led to their creation. To get good at predicting words requires a detailed world model. LLMs *demonstrably* have *immensely* detailed world models (as demonstrated by probe studies). Heck, even seq2seq has a world model.
September 24, 2024 at 3:14 PM
Hi @fchollet.bsky.social! 👋

I have made a small improvement in your "character-level seq2seq translation" example from #Keras documentation. I have added attention layers and some custom metrics

I hope it would be helpful to add more examples to the documentation ☺️

gist.github.com/DelgadoPanad...
Keras Character-level recurrent sequence-to-sequence attention model
Keras Character-level recurrent sequence-to-sequence attention model - Character_level_recurrent_sequence_to_sequence_attention_model.ipynb
gist.github.com
January 9, 2025 at 9:31 AM